EAvatar: Expression-Aware Head Avatar Reconstruction with Generative Geometry Priors
Authors: Shikun Zhang, Cunjian Chen, Yiqun Wang, Yong Li
Organizations: Department of Data Science and AI, Monash University, Melbourne, Victoria, Australia · College of Computer Science, Chongqing University, Chongqing, China
High-fidelity head avatar reconstruction plays a crucial role in AR/VR, gaming, and multimedia content creation. Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated effectiveness in modeling complex geometry with real-time rendering capability and are now widely used in high-fidelity head avatar reconstruction tasks. However, existing 3DGS-based methods still face significant challenges in capturing fine-grained facial expressions and preserving local texture continuity, especially in highly deformable regions. To mitigate these limitations, we propose a novel 3DGS-based framework termed EAvatar for head reconstruction that is both expression-aware and deformation-aware. Our method introduces a sparse expression control mechanism, where a small number of key Gaussians are used to influence the deformation of their neighboring Gaussians, enabling accurate modeling of local deformations and fine-scale texture transitions. Furthermore, we leverage high-quality 3D priors from pretrained generative models to provide a more reliable facial geometry, offering structural guidance that improves convergence stability and shape accuracy during training. Experimental results demonstrate that our method produces more accurate and visually coherent head reconstructions with improved expression controllability and detail fidelity.
Figures & tables
Figure 1. Qualitative comparisons of different methods on the cross-identity reenactment task. Each column represents a different method, and each row shows results at consecutive time steps in the reenactment sequence. From left to right: NeRFace, HAvatar, GHA, GaussianAvatars (GA), TensorialGA, and Ours. Our method more faithfully transfers expressions and maintains identity across frames, producing photo-realistic results.
Figure 2. Overview of the proposed two-stage framework. In Stage 1 (Sec. 3.3), multi-view images are first processed by a 3DMM estimator to provide coarse geometry, pose, and expression cues. Based on these cues, the framework performs SDF-based geometry modeling, extracts a surface through differentiable surface extraction, aligns the resulting mesh with an identity-aware prior mesh, and then initializes the dynamic Gaussian avatar. In Stage 2 (Sec. 3.1, Sec. 3.2, and Sec. 3.4), the initialized avatar is driven by pose and expression signals to predict dynamic offsets and attribute variations, which are further refined in locally highly deformable regions through the proposed control Gaussian mechanism and Gaussian splitting, followed by differentiable rendering and reconstruction optimization to produce the final dynamic head avatar.
Figure 3. Qualitative comparisons of different methods on self-reenactment task. From left to right: NeRFace, HAvatar, GHA, GaussianAvatars(GA), TensorialGA and Ours. Our method can reconstruct details like eyes, teeth, etc. with high quality.
Figure 4. Another cross-identity example evaluated under the same setting as Fig. 1 . Columns correspond to methods and rows to consecutive frames. Our approach continues to exhibit consistent identity preservation and expression fidelity over time.
Method
Case 1
Case 2
Case 3
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
NeRFace
21.38
0.746
0.283
21.08
0.847
0.279
20.19
0.817
0.223
HAvatar
22.62
0.802
0.247
21.63
0.861
0.264
22.16
0.878
0.184
GHA
24.02
0.814
0.203
25.47
0.879
0.147
24.56
0.902
0.144
GaussianAvatars
21.61
0.796
0.237
25.73
0.850
0.151
21.23
0.846
0.197
TensorialGA
23.23
0.807
0.213
24.29
0.860
0.242
23.60
0.895
0.162
Table 1. Quantitative evaluation results of NeRFace, HAvatar, GHA, GaussianAvatars, TensorialGA and our full method on the self-reenactment task. ↓ indicates lower is better, ↑ indicates higher is better.
Figure 5. Results of our expression-controllable head avatar generation. As demonstrated, the generated results accurately capture facial expressions while preserving fine-grained local details.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
NeRFace
20.43
0.842
0.215
HAvatar
21.16
0.878
0.184
GHA
22.89
0.880
0.167
GaussianAvatars
19.83
0.844
0.217
TensorialGA
21.51
0.876
0.182
Ours
23.22
0.882
0.165
Table 2. Quantitative evaluation of our method and other SOTA methods on 3D consistency for novel view synthesis.
Figure 6. Qualitative comparison with prior methods on the novel view synthesis task. We use 8-view synchronized videos for training the avatar and the remaining 8 new views were used to test.
Figure 7. Ablation study of the structure-aware modeling strategy: by introducing an explicit mesh as a shape constraint, our strategy ensures that better contour and pentagonal shapes are obtained in the first stage.
Method
Case 1
Case 2
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Baseline
23.28
0.812
0.217
25.47
0.879
0.147
+control
23.36
0.816
0.214
25.55
0.879
0.147
+control & split
23.47
0.818
0.199
25.74
0.880
0.145
+control & split & structure: (Ours)
23.66
0.822
0.196
26.72
0.884
0.141
Table 3. Impact of the different modules proposed in our approach on the self-reenactment task.
Figure 8. Ablation study on controllable Gaussian mechanism. The figure shows the results of self-reenactment task.
Figure 9. Ablation study of the controllable Gaussian mechanism on the cross-identity reenactment task.
Figure 10. Ablation study of the split strategy on the cross-identity reenactment task.
Figure 11. Visual comparison of Stage I loss ablations after the same 10 training epochs. From left to right: ground truth, full model, w/o {Loffset,Llap} , w/o Llmk , and w/o Llap .
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Ours
24.82
0.903
0.141
w/o offset+lap
24.67
0.896
0.145
w/o landmark
24.71
0.902
0.144
w/o lap
24.78
0.900
0.142
Table 4. Quantitative comparison of different Stage I loss ablations on the self-reenactment task (Case 1). The metrics are reported on the final reconstruction after completing Stage II.
Threshold
Case 1
Case 2
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
0.15
23.86
0.795
24.72
0.858
0.30
24.24
0.814
24.82
0.903
0.40
24.01
0.796
24.48
0.901
Table 5. Impact of different control thresholds on two representative cases.
Figure 12. Ablation study on splitting threshold. The figure shows the results of self-reenactment task.
Method
Training Time ↓
Rendering FPS ↑
NeRFace
4 days
˜ 0.06
HAvatar
7 days
˜ 7
GHA
3 days
˜ 32
GaussianAvatars
1 day
˜ 183
TensorialGA
1 day
˜ 150
Ours
2.5 days
˜ 32
Table 6. Comparison of training time and rendering speed across methods.
Method
Prep. (h) ↓
FPS ↑
PSNR ↑
SSIM ↑
LPIPS ↓
GaussianAvatars
32.7
183
22.50
0.826
0.207
Ours
12.0
32
25.07
0.857
0.168
Table 7. Efficiency–performance trade-off comparison with GaussianAvatars. Results are averaged across subjects under our more challenging evaluation setting.
Figure 13. Visualization of the selected Control Gaussians for different subjects. Each row shows two representative frames with clear expression variation and the corresponding 3D distribution of the selected Control Gaussians. The highlighted points consistently concentrate in semantically expressive and highly deformable facial regions, while still showing subject-specific variation.
3D Gaussian Splatting (3DGS) provides an efficient method for high-quality scene reconstruction using anisotropic Gaussians. Recently, 3DGS-based methods have significantly improved the rendering quality of human avatars while enabling real-time performance. However, existing methods suffer from a magnitude mismatch in the number of Gaussians generated by image-based and 3DMM-based approaches. This discrepancy results in reconstructed expressions that lack fine-grained detail. In this paper, we introduce a novel method for reconstructing an animatable head avatar from a single image. We propose a Graph splitting network to progressively generate Gaussians from coarse to fine using an autoregressive architecture. To address the graph inconsistency caused by split Gaussians, we employ a mesh topology extension method to align the GNN's connectivity with the increased Gaussian count. Furthermore, we introduce a novel density control method that includes a gating mechanism that generates soft masks for Gaussians, preventing over-densification after the splitting operation. This allows for dynamic control over Gaussian density across different facial regions. For smooth and rapid training, we employ a delayed filtering strategy to avoid re-computing the graph topology during training. Experimental results demonstrate that our autoregressive structure effectively improves expression representation ability by progressively splitting Gaussians. This process, enabled by the GNN-guided splitting, synthesizes more precise facial details and achieves higher reconstruction quality.
Hongzhe Liao, Chuhua Xian, Hongmin Cai +2
South China University of Technology · University of Tokyo · The Hong Kong University of Science and Technology
We propose S-Avatar, a novel method for generating photorealistic 3D head avatars from a single image using a diffusion-guided 3D model generation module and strategies for animating 3D Gaussian Splatting (3DGS). While single-image head avatar reconstruction is crucial for lifelike Virtual Reality (VR) applications, existing approaches often struggle to preserve 3D consistency under unseen viewpoints. S-Avatar addresses this limitation through a three-stage pipeline. First, a high-resolution 3DGS is synthesized directly from a single image using a diffusion-based Gaussian splat generation module. Next, the parametric head model FLAME is aligned with the generated 3DGS by optimizing its parameters and spatial transformations. Finally, to adapt the 3DGS to FLAME variations, we construct a binding template that encodes the spatial relationship between the initial splats and FLAME. The dynamic 3D head avatar can then be rendered in real time by deforming the 3DGS with the binding template. By combining diffusion-guided canonical 3DGS generation with FLAME-based control, our method achieves efficient and accurate reconstruction with enhanced 3D consistency. Evaluations on public datasets demonstrate that S-Avatar outperforms state-of-the-art methods in novel-view and expression generation, achieving superior realism and consistency. Consequently, our approach represents a significant advance in accessible avatar creation, applicable to a wide range of VR/AR applications. The project page is available at https://github.com/hailsong/savatar.
Reconstructing photorealistic and animatable 4D head avatars from a single portrait image remains a fundamental challenge in computer vision. While diffusion models have enabled remarkable progress in image and video generation for avatar reconstruction, existing methods primarily rely on 2D priors and struggle to achieve consistent 3D geometry. We propose a novel framework that leverages geometry-aware diffusion to learn strong geometry priors for high-fidelity head avatar reconstruction. Our approach jointly synthesizes portrait images and corresponding surface normals, while a pose-free expression encoder captures implicit expression representations. Both synthesized images and expression latents are incorporated into 3D Gaussian-based avatars, enabling photorealistic rendering with accurate geometry. Extensive experiments demonstrate that our method substantially outperforms state-of-the-art approaches in visual quality, expression fidelity, and cross-identity generalization, while supporting real-time rendering.